Blind Source Separation of Ongoing Time Series Data

Question

I have a blind source separation problem that matches the form of "isolating conversations in a crowded room" and have been looking into ICA as an option. Basically, a sensor array will have overlap in its readings that need to be factored out. However, I need to be able to do this in an online learning context where the solution is based on the aggregated Time Series data to date. I was wondering if ICA is generally an appropriate solution or if I am approaching this correctly. Any suggestions on direction would be appreciated.

I am not entirely sure of the terminology I should be using so please bear with me.

Answer 1

Online versions of ICA do exist. ICA is appropriate if its assumptions are met: Sources must be non-Gaussian. Measurements must be linear combinations of sources. The mixing coefficients relating sources to measurements must not change over time (but some ICA variants claim to handle nonstationary/time-varying mixtures). The number of sources must not exceed the number of sensors. ICA assumes they're equal, but often returns reasonable results when the number of sources is fewer than the number of sensors. Reducing the dimensionality with PCA prior to ICA may also help in this situation (and can also be performed online).